Few-Shot Learning (少样本学习)

Few-Shot Learning (FSL) is a specialized subfield of machine learning (ML) designed to train models to recognize and classify new concepts using a very small number of labeled examples. In traditional deep learning (DL), achieving high accuracy typically requires massive datasets containing thousands of images per category. However, FSL mimics the human cognitive ability to generalize rapidly from limited experience—much like a child can recognize a giraffe after seeing just one or two pictures in a book. This capability is essential for deploying artificial intelligence (AI) in scenarios where collecting large amounts of training data is prohibitively expensive, time-consuming, or practically impossible.

少量学习的核心机制

FSL的主要目标是通过利用先验知识来减少对大规模数据收集的依赖。 该模型无需从零开始学习模式，而是利用包含少量标注样本的"支持集"来理解新类别。这通常通过元学习（亦称"学习如何学习"）等先进技术实现。在此范式下，模型通过训练多种任务来习得最优初始化或更新规则，从而能以最小调整适应新任务。

另一种常见方法是基于度量的学习，模型通过嵌入技术将输入数据映射到向量空间。 在此空间中，相似元素 被聚类至邻近位置，而相异元素则被推离。诸如 原型网络等算法会为每类计算均值表示（即原型）， 并根据classify 查询样本与这些原型的距离classify 。该方法通常依赖于 在大型通用数据集预训练阶段 培养的特征提取能力。

实际应用

少样本学习正在改变那些因数据稀缺而阻碍人工智能技术应用的行业。

医学影像与诊断

In the field of medical image analysis, obtaining thousands of labeled scans for rare pathologies is often unfeasible. FSL allows researchers to train computer vision (CV) systems to detect rare tumor types or specific genetic anomalies using only a handful of annotated case studies. This capability democratizes access to advanced diagnostic tools, a goal pursued by institutions like Stanford Medicine, helping to identify conditions that would otherwise require specialized human expertise.

工业质量控制

现代制造业人工智能高度依赖自动化检测。然而某些特定缺陷可能极为罕见，导致难以构建大规模的"不良"零件数据集。FSL技术使异常检测系统仅凭少量图像即可学习新型缺陷特征，使工厂操作员能够快速更新质量保证协议，无需停产收集数据，从而显著提升动态生产环境中的效率。

区分相关概念

区分FSL与类似的低数据学习范式有助于理解其特定的应用领域：

• 迁移学习：自监督学习（FSL）通常被实现为迁移学习的一种特殊极端形式。标准迁移学习可能通过数百张图像对YOLO26等模型进行微调，而FSL则针对每类仅有5至10张图像的场景（称为"N类K样本"分类）。
• 单样本学习：这是 自由样本学习（FSL）的严格子集，要求模型仅从单个标注样本中学习。该技术常用于 人脸识别领域，通过比对存储的单张照片来 验证身份。
• 零样本学习：不同于需要至少少量视觉支持集的监督学习，零样本学习在训练过程中无需目标类的视觉样本。它依赖语义描述或属性（如文本提示）来识别未见过的对象。

Ultralytics的实际应用

In practice, one of the most effective ways to perform Few-Shot Learning is to leverage a highly robust pre-trained model. State-of-the-art models like the newer YOLO26 have learned rich feature representations from massive datasets like COCO or ImageNet. By fine-tuning these models on a tiny custom dataset, they can adapt to new tasks with remarkable speed and accuracy.

以下Python 演示了如何使用小数据集训练模型： ultralytics 包，有效实现少样本适应：

from ultralytics import YOLO

# Load a pre-trained YOLO26 model (incorporates learned features)
model = YOLO("yolo26n.pt")

# Fine-tune on a tiny dataset (e.g., coco8 has only 4 images per batch)
# This leverages the model's prior knowledge for the new task
results = model.train(data="coco8.yaml", epochs=20, imgsz=640)

# The model adapts to detect objects in the small dataset
print("Few-shot adaptation complete.")

挑战与未来展望

While powerful, FSL faces challenges regarding reliability. If the few provided examples are outliers or noisy, the model's performance can degrade, a problem known as overfitting. Research into data augmentation and synthetic data generation is critical for mitigating these risks. As foundation models become larger and more capable, and tools like the Ultralytics Platform simplify model training and management, the ability to create custom AI solutions with minimal data will become increasingly accessible to developers worldwide.